Abstract
The rapidly improving capabilities of large language models (LLMs) create both opportunities and threats for professions. Organizational scholars have been quick to raise the alarm about the perils that generative artificial intelligence (AI) poses for society and academic scholarship in general (Lindebaum et al., 2025) and qualitative research in particular (Nguyen & Welch, 2026). Any technology that is as powerful as LLMs inevitably will create unintended and unforeseen negative consequences; for example, chatbots allow users to rapidly create shallow qualitative analyses, made-up quotes, and bogus theory. However, critics often ignore a central argument in technology studies: The impact of any technology depends on the social practices in which it is embedded.
What is our alternative, really? Knowledge work is rapidly changing across professional groups. Lawyers, physicians, and engineers now increasingly use generative AI to improve the quality and speed of their work. Can management scholars decide to sit this out? Proponents of an artisanal anti-AI culture may seek to stigmatize the use of LLM-based tools to aid research, discouraging and punishing AI use. This would leave our profession, tasked to educate future managers, ill-prepared to understand AI-augmented knowledge work. Because craft and community are invaluable to our profession (Bechky & Davis, 2025), we should engage in practice development around AI or we risk becoming obsolete in the eyes of external stakeholders on whom we depend.
At best, AI provides qualitative researchers with a pedagogic opportunity in the same way that business analytics provided earlier for statisticians. Our expertise can inform “lay qualitative research” conducted across the private and public sectors by millions of white-collar workers. In this essay, I will focus solely on qualitative analysis, although AI can support many parts of the research process.
LLMs in Qualitative Analysis
LLMs are based on the transformer algorithm that effectively captures complex interaction patterns in data, such as semantic meanings in sentences and paragraphs. The technology became widely known with ChatGPT, and many still view “GenAI models . . . as a type of chatbot” (Nguyen & Welch, 2026, p. 4). More recently, however, chatbots have been complemented with software-driven workflows and agents.
Chatbots
Although chatbots can identify quotes and form categories from qualitative data, the enthusiasm concerning generic chatbots is not wholly warranted. Most significantly, chatbots involve a tradeoff between computation and response time. Because chatbots are optimized to provide rapid answers, and perhaps because the extent of computation influences costs, these services are designed to process only a fraction of the provided text with the AI model. Services such as ChatGPT apply opaque search algorithms to identify the most relevant passages relating to the user prompts that are passed to the model to form the response. Such functionality excels in answering specific user questions but does not constitute a sound basis for systematic and thorough analysis.
AI Workflows
To analyze empirical data more systematically and counter so-called hallucinations, the creation of inappropriate but plausible-sounding text, diverse software tools have been developed to “harness” LLMs (see, e.g., libraries mall and elmer for the statistical software R). AI-native analysis tools and legacy qualitative analysis software packages offer AI functionalities to systematically code and categorize content with predefined workflows. This software layer can instruct the LLM to extract specific quotes from documents and search that they actually exist. The greatest benefit of these workflows lies in their mechanistic nature: Researchers can offload some of the noncreative tasks, such as identifying specific kinds of passages or categorizing quotes in descriptive categories. This allows the researcher to focus on more creative work: deciding on theoretical framing and the kinds of quotes to extract, inventing new ways to categorize data, and crafting explanations for relationships between categories.
Agentic AI
The most recent incarnation of AI, agentic systems, represent a powerful use of LLMs where processes involving dozens of sequential steps are autonomously defined and executed by the AI. Agentic approaches are the most mercurial because they abandon the mechanistic predefined nature of AI workflows and rely on the growing capabilities of the models themselves to plan, orchestrate, and monitor complex tasks. Although agentic AI capabilities have improved greatly, they remain inherently erratic.
Agentic AI can competently do many qualitative research tasks, such as conducting open-ended exploration of qualitative datasets, developing code books, organizing evidence into tables, and even proposing causal boxes-and-arrows process models (see my experiment with U.S. presidential transition interviews at https://github.com/hschildt/agentic-qual). The resulting figures and tables have the outward resemblance of an experienced researcher. However, as I scrutinized the evidence more closely, it became clear to me that the model was very weakly grounded: The agentic process had force-fitted data into the interpretation I had asked for and relied on a single interview for many of the key observations, and somewhat ill-fitting quotes were chosen to stand in for concepts they didn’t really match. Despite these limitations, AI outputs can generate useful ideas that help researchers assess and refine their research question and tables of quotes that can further spur new ideas. Ultimately, I consider this the most appropriate goal of AI use in an academic context: not to replace human ingenuity but to provide ideas and evidence as raw materials for it.
Why should we embrace these AI tools, then? By simply speeding up the coding process, a researcher can investigate more alternative framings and coding structures and avoid becoming “locked in” to their chosen framing. Moreover, AI can help us be more rigorous in how we treat data. Instead of relying on “star informers.” AI helps comb through every archival document and interview to surface supporting and contradicting evidence for our arguments. Finally, we are likely to see new strands of qualitative research emerging because AI allows us to use larger datasets than previously, develop new methods, and address questions we could not before (Gartenberg et al., 2026).
Developing New Practices: Two-Way End-to-End Transparency
Responsible AI use requires oversight. Because LLMs are based on training data corpora and human reinforcement learning, their ability to interpret text varies across contexts. Research also has shown that LLMs exhibit strong biases that even exceed their training data (Hofmann et al., 2024). Luckily, researchers have always had to evaluate interpretations; no human coauthor has ever been devoid of biases. As with humans, addressing biases requires transparency. Traditionally, researchers have produced tables of quotes and laid out the evidence for their coauthors and journal reviewers. With AI, the practice should be no different: Scrutinize the interview context and extensive quotes to judge AI outputs.
Our tools should be designed for what I call two-way end-to-end transparency. First, researchers ought to see, with minimal effort, the original documents from which AI formed its interpretations. Second, researchers should be able to go through the original data and see which key parts of text were picked up by AI and which were not, and to what effect, constituting two-way transparency. As agentic AI becomes increasingly powerful, we need to design AI solutions to visualize and help track the steps agents took, such as which phases AI used to search through empirical corpus, which document segments were scrutinized, how the categories were formed and adjusted, and which approaches were tested and then abandoned by the AI. This constitutes end-to-end transparency for complex AI processing.
Cultural Impact of AI: Building a Stronger Community
The development of shared norms and practices for AI use will require collective work from our profession. More broadly, many scholars now see the risk of LLMs to erode our professional standards (Bechky & Davis, 2025). Some have suggested that once professionals use AI to generate any content, over time they will cease to do it themselves (Lindebaum et al., 2025). Professions and professional organizations have always been able to assess knowledge and effectively reward and sanction individuals based on the quality of their outputs. Despite the early negative impact of AI on research due to its superficial use in writing (Gartenberg et al., 2026), AI technology does not need to dissolve professional norms and safeguards.
A more serious concern, voiced in many seminars, is the effect that AI has on junior academics. Growing pressure created by performance metrics has increased the likelihood that LLMs will erode intellectual engagement (Bechky & Davis, 2025). The lure of automation may prevent juniors from gaining the experience, and the lack of experience may prevent them from evaluating AI outputs. These challenges posed by AI are shared across professions.
My recommendation is to reinforce professional norms and motivate scholars to engage more deeply with both the empirical phenomena and the ideas of existing literature. AI can further facilitate this by searching and “translating” knowledge in a form that is easier to grasp. Humans must take responsibility for knowledge production, requiring scrutiny of the way AI augments our work processes. The role of senior scholars thus will be to kindle curiosity in junior scholars.
AI and the Future of Business Schools
The use of AI in any professional context can be a thorny issue because it touches the core of our identities: what we do and how we work. This makes AI a polarizing topic, with “conservative” voices warning about the corrosive effect on qualitative research (Nguyen & Welch, 2026) and business education (Lindebaum et al., 2025). I propose a competing “progressive” vision, where our efforts to leverage AI in qualitative research help us pioneer our new responsible practices and position as authorities on the impact of AI on knowledge work.
By developing innovative uses for AI, we can become attuned to both the risks of AI and practices that mitigate those risks. Such an attitude was illustrated by a medical doctor attending my executive MBA course, who noted that the ability of LLMs to diagnose diseases is already so good that it may be unethical for physicians not to consult AI as a complement to their own expertise. Perhaps soon we will consider it unethical not to use AI in qualitative research to assess alternative explanations and contradictory evidence?
Qualitative research has much in common with professional work: analysis that dissects phenomena into “constitutive parts” and synthesis that composes a coherent account from those parts. Lawyers, public servants, and management consultants all engage in “lay qualitative research” as they group observations together and then link them to diagnoses or narratives. In many ways, business schools are ideally positioned to pioneer new knowledge work practices that go beyond our own field. We can ground those practices in AI-enabled qualitative research and PhD training, empirical research on AI use in business, and the new pedagogic practices that incorporate AI. The financial interest of large technology companies may be aligned with the automation rather than augmentation of expert work, whereas business schools are focused on empowering their students.
Conclusion
We have much to gain from experimenting with AI in qualitative analysis and developing responsible and transparent research practices. If we ignore AI, we risk irrelevance. By developing responsible qualitative research practices, we can help companies leverage knowledge embedded in textual data, replicating the success of business schools in introducing data science to managers over previous decades.
Footnotes
Acknowledgements
I am grateful for the excellent advice I received from John Amis and Tammar Zilber throughout the writing of this essay. My thinking about qualitative research and the role AI can play in it has been shaped by continuous conversations with Stine Grodal, Saku Mantere, and Farah Kodeih. Claude (Anthropic) helped me eliminate annoying grammatical errors and repetitiveness.
Declaration of conflicting interest
The authors declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: I am a co-founder of Skimle.com, an AI analysis platform.
